Removing bias in answers about guardrails HOT 2 CLOSED

Thanks for testing out the library @bluewin4!

This is a pretty interesting suggestion. I think this could be pretty useful. My main concern would be around managing the different prompts for the two LLM API calls. The first prompt should ideally contain the full information around what the content of the answer should be, and the second prompt should contain info about the structure that the content should be formatted into.

One way to implement this could be to have the user continue to write a single rail spec, and then Guardrails compiles the rail spec into the prompt needed for each LLM API call.

Another thing to figure out would be reasking -- reasking would also need to be broken out into reasking for content and reasking for structure.

I'll dig deeper into this and see what the MVP is that I can implement. If you have other suggestions, lmk!

Also, we chatted about this briefly in the discord in case you're interested in following along.

from guardrails.

I've also found it useful sometimes to invoke a LLM call first, then run guardrails over its result. But for now we've decided to leave this to the user, not as an internal part of the library.

from guardrails.